Advanced Recognition of License Plates (ARLP)
Advanced Recognition of License Plates (ARLP) is a neural network-based system for Indian vehicle license plate detection and recognition.
Created on 27th October 2024
•
Advanced Recognition of License Plates (ARLP)
Advanced Recognition of License Plates (ARLP) is a neural network-based system for Indian vehicle license plate detection and recognition.
The problem Advanced Recognition of License Plates (ARLP) solves
The Advanced Recognition of License Plates (ARLP) project aims to accurately recognize and interpret license plates on Indian vehicles, with applications in:
- Traffic Management: Automatically identifying and recording vehicle plates to improve monitoring and traffic management.
- Law Enforcement: Aiding in the tracking of stolen vehicles and identifying those involved in criminal activities.
- Parking Management: Automating vehicle entry and exit, enhancing security and reducing manual checks.
- Toll Collection: Improving toll efficiency by recognizing plates and reducing wait times.
- Access Control: Ensuring secure, automated entry to restricted areas by verifying plates.
Using advanced technologies like CNNs, Inception ResNet v2, and attention mechanisms, the ARLP system enhances accuracy, efficiency, and automation in license plate recognition, promoting overall safety and convenience.
Challenges I ran into
In developing the Advanced Recognition of License Plates (ARLP) project, several challenges arose:
-
Custom CNN Model Accuracy:
- Achieving sufficient accuracy with the Custom CNN model was challenging, as initial accuracy reached only 64.5%, indicating a need for extensive hyperparameter tuning.
-
Register Spillage in Local Memory:
- Training was impacted by register spillage into local memory, as indicated by the warning:
ptxas warning : Registers are spilled to local memory in function 'loop_add_subtract_fusion_49', 224 bytes spill stores, 224 bytes spill loads
- This required optimization of the code and model architecture.
- Training was impacted by register spillage into local memory, as indicated by the warning:
-
Hyperparameter Tuning:
- Extensive tuning was necessary to enhance the Custom CNN model's accuracy, requiring experimentation with different configurations.
-
Comparative Performance:
- The Custom CNN model initially struggled to accurately locate plates, while Inception ResNet v2 achieved 94.3% accuracy, necessitating a balance between custom model refinement and transfer learning.
To address these, we refined the architecture, optimized hyperparameters, and utilized Inception ResNet v2's strengths to improve accuracy and robustness.
Technologies used